Eight Years of GraphQL
magiroux.com
magiroux.com
Who would torture themselves like that?
Isn't the whole point that your frontend can make the exact queries it needs and load the exact data it needs?
Namely, last I checked, client libraries for working with graphql are only good with JS. I tried working with graphql a few years ago in python and the only two client libraries absolutely sucked. Server libraries were great, but python clients sucked, badly. I ended up writing bare requests with a hardcoded heredoc for the query and endless square brackets to get the fields for the little data I needed.
Maybe the situation improved dramatically in the last three years, but I can't imagine so dramatically.
I wouldn't pick graphql as a private backend API in a million years. Well, maybe if ever single service was written in nodejs with no possibility of using other languages.
As for language support, I’ve used a couple of libraries on both the client and server side for both Go and PHP, and they seem pretty good. JavaScript/TypeScript definitely seem to be the first-class citizens though.
In your case of many disparate services, the nightmare is just the same with REST.
For GraphQL, you know it will be a nightmare. For REST, you don't, because you don't know what the actual API is. It could be a stopwatch API, obviously a stopwatch REST API is not comparable to GraphQL.
It's literally meant to be a gateway to a bunch of different external systems working in federation.
It's the exact same scope if you used a RESTful graph based data access API that connected to external systems..
Outside all the GraphQL jargon and gobbledygook you'll see that GraphQL is not some groundbreaking, unique or foundational product... The concept of generating graph based data access APIs from schemas is probably older than the concept of phones that can send emails. The only "innovation" is the "query language"...
> That doesn't matter at all.
I honestly don't know why I'm taking this seriously. Yes, it matters, your entire point was a comparison.. that your comparison is foundationally invalid matters when assessing the validity of your entire argument..
Hence your comparison is invalid
> you don't have any guarantees about the API and thus it's a much harder target to support.
You won't find a developer worth hiring who thinks OpenAPIs are more difficult to work with than GraphQL.
Every REST API can be an OpenAPI
Practically every popular general purpose language has an OpenAPI client gen implementation.
And supporting a GraphQL API is not easier than supporting a REST API. All I have to do for my stopwatch API is handle one route.
> You're limiting it to json:api which is nice
Certain constraints do come with valid comparisons, comparing like things is a requirement of a valid pros/cons comparison, not a niceity.
It makes no sense to explain the pros of using one thing over an other when no one is thinking about using one thing as a replacement for another.
It's like talking about the pros/cons of using your proprietary web based chat platform vs. using UDP. Not using UDP for the chat or anything. Just comparing the chat with UDP itself, the whole protocol. No one considering your chat is going to step back and go "wait why am I using a web chat when I can just use UDP?"
> I have never even seen a company use that
My point isn't that a company does/doesn't use it. I mentioned JSON:API because it's a valid basis of comparison, no other reason.
> What happens in reality is that I get 20 apis from 20 teams, each of them completely different principle - and am supposed to build an universal auth layer for that. That's much more painful than in the companies that decided to go for GraphQL only.
But you have by your own admission never seen another company use JSON:API.. therefore you don't actually know that it is more painful than it would be for all REST APIs.
For all I know you could, again, be talking about my Stopwatch REST API... hence your comparison is invalid. By the way, auth is easier for my stopwatch API than it is for GraphQL. :)
Which a REST API can do.
This argument is quite a goalpost shift from your original point: that it is easier to implement authz/access control in GraphQL than for a REST API.
An API to generate PDFs is not comparable to GraphQL, because it's not a Graph based data access API...
You're so disingenuous and unserious it's insane. This is why people hate GraphQL and it's devs: 100% of your talking points are non-sequitur copy paste arguments. It doesn't matter how much context is provided to prove what you're saying is nonsense — you'll just bulldoze right along with said nonsense anyways. It's like I'm talking with a zombie.
For the purpose of accountability I'm pointing out that so far you've:
- made an invalid comparison
- tried to elaborate on that comparison only to make several other logical errors
- goal post shifted to a totally different point than your original point after being called out
you've already pointed out that you have zero experience with JSON:API and that you can't come up with anything good to compare GraphQL with sort of demonstrates that your reasoning behind using it is not based on anything other than marketing as you have never actually used an alternative to GraphQL... Everything you've said I've only ever seen in GraphQL promo material.
So basically you got shilled and turned into a shiller.
However, for all technical things, then i've noticed that things can be made quite obvious if you really focus on the exact problem you actually have to solve, and have a good understanding of its nature.
It's not really torture. I do. I have built a private federation API in a system where there's a "mothership" service that needs to talk to individual websites.
Those websites are WordPress at the moment but they may be Magento, PrestaShop or whatever in the future.
GraphQL means I can template the API calls used to keep them in sync with the mothership. It's awesome.
I also use it for the frontend/backend connection of a couple of admin APIs and an intranet app.
Nuxt frontends, Laravel/Lighthouse headless backends. The only pain point is the unnecessarily complex Apollo stuff in the client, but that gives you really nice smart queries in Nuxt 2; it may not be that crucial in Vue 3/Nuxt 3.
It's even better with Vue 3 (or Vue 2 with the composition API anyway), especially when you combine it with graphql-codegen. Smart queries become just another composable, no `this.$apollo` to be seen. Got that going with API Platform on the backend myself, which has really nice integration with symfony, though it speaks only a bare-bones dialect of graphql as opposed to Lighthouse with all its crazy directives. No persisted queries either, but it's an internal app with bandwidth to burn.
Been eying wp-graphql for my one wordpress project. From my dabbling with it, the DX feels a lot nicer than the WP REST api, though I'm sure you know that's an awfully low bar to clear.
I did not know about this, thank you.
> Been eying wp-graphql for my one wordpress project. From my dabbling with it, the DX feels a lot nicer than the WP REST api, though I'm sure you know that's an awfully low bar to clear.
Yes on both scores. There's a reasonably useful Woocommerce binding:
https://github.com/wp-graphql/wp-graphql-woocommerce
Which is what I was using for the federated storefronts, with the API locked down so it could only be accessed from a back-office client I hooked up with a Laravel worker. At the time it was very early days, but wp-graphql has quite a sane interaction with the hooks/actions model, so it's not awfully difficult to add/patch/override the things you need.
For backend services in particular literally anything else is probably a better choice - rest, grpc, soap, a socket. I honestly don't know why more people don't give twirp a shot.
Oh, and we get DB schema to frontend propagation of changes to the type.
This is by far the best experience I have had for private APIs - Where we don't need to maintain backwards compatability.
I do get that the HN crowd is super adverse to graphql, but I fail to see other arguments than matters of taste.
In a setup where you can deprecate fields and remove them after X weeks I think it could wor well - My impression is just not that enterprise works like that?
I would probably go the route of having fully versioned APIs in the specific setup (GraphQl or not) - which is something that would be a hassle with a tight coupling between the Prisma schema and the GraphQl schema as proposed.
Thinking about it, all the situations where I have successfully used GraphQL, there has been a relatively tight coupling between the database schema and the graphQl schema. Previously it has been Ecto + Absinthe + Zeus (Elixir on the backend).
You’re right that there will need to be long-term support of API versions (or at least the current one and maybe the previous one).
My instinct is to give the users the full objects and let them select the fields they want on their side. It’s less work on our side. Also I’m not sure about customer comfort with GraphQL.
Most interesting applications have some type of properties of entities that transcent a static object.
Thanks for sharing feTS. It looks pretty awesome and I will be checking it out.
Second, I don't think it's very comparable. The big drawback of "RESTful APIs" is that you cannot combine things. You call an endpoint and you add some query parameters, that's pretty much it.
In Graphql, you can combine and even nest queries. You simply cannot (or don't want to) generate all combinations in advance, so you'll decide adhoc in your code. Therefore you need a library that can do these combinations adhoc for you in a typesafe way.
Okay, so write fragments. However, without a framework like Relay (and soon Apollo), you still receive, at runtime, the entire network response. ie there is no masking at the fragment level. This means that there still is an implicit dependency between components, and removing fields is still dangerous.
That’s why you need Relay.
I make that case here - https://youtu.be/lhVGdErZuN4?si=tOJNWl-j-Uj28RUs
GraphQL really shines with data intensive applications where you need pagination, filtering, sorting, projections etc.
The Graphql server itself is still publicly exposed to the internet, but the ability to query is not. Queries have to be whitelisted ahead of time (persisted queries).
Instead they now focus on perf problems of downstream federated GraphQL queries. And query complexity. And unbounded queries. And extreme overfetching of data that all clients inevitably do. And...
Unbounded queries and extreme overfetching in general are problems that are ... easy-ish to solve when it's totally internal. Just be measuring perf, tag queries you are making to the frontend page you're making them from, and coordinate with frontend people if there's a problem.
Performance doesn't magically improve if you're using REST. And joins don't magically make things slower than "join via HTTP request" either. There might be patterns that are dangerous in GraphQL but honestly I feel like most internal APIs would benefit from easier scoping of fields and the like (rather than everyone re-inventing ad-hoc expansion and filtering)
In REST you know the exact query and can optimise the hell out of it. Not so in GraphQL, where each query is potentially a new, never before seen request. Unless you use and optimise for persisted queries which just makes it REST with extra steps. And for large companies even that may not be an option since every query will end up being a persisted query once you deploy to production.
> Unless you use and optimise for persisted queries which just makes it REST with extra steps.
GraphQL offering a syntax and a way to state exactly what will be used makes it possible to chip away what is sent over the wire in persisted queries. This is not a negligable thing in environments with very wide and deep data models. These are places where their "REST" queries end up with "GraphQL with extra steps" through path inclusion/exclusion logic and expansion logic.
Obviously at the end of the day context matters most, but there's been loads of places and APIs I've used where you could feel REST causing performance issues in an end-to-end way, either by generating N+1 HTTP queries, or just shipping way too much data to show a list of names of resources.
Because people love chasing tech fads.
Here’s my workflow for creating an API with Postgraphile:
create view graphql.object as select some,columns from table;
It’s trivial to query it with curl, I’d give an example but I’m afk rn.
I’ve been using GraphQL for about the same amount of time as in the article and it solved a bunch of problems for me.
It’s so easy to use, and saves so much time - once you spend the time to understand it.
If you know Postgres, I would recommend taking a look at Postgraphile. It’s awesome, and comes with an explorer web UI that really helps (GraphIQL with extras). Everything happens in real time. so if you update a view, the UI updates.
There are lots of GraphQL clients but many of them do all sorts of crap you don’t need. I just use graphql-request which is super simple. But of course you can just use fetch() too.
There are also lots of “standards” for GraphQL that make it seem more complex than it is. Ignore that stuff and just start playing with a good server like Postgraphile.
Good luck!
The complexity of GraphQL in fact begins there, and also sort of explains a lot of why GraphQL is all but simple: Why am I using a query language instead of just passing an AST via JSON, a data format every general purpose language supports very well these days?
The answer to the above question, and most of GraphQLs other complexities: Some arbitrary design decision.
Another example: GraphQL could've easily been expressed as a REST API, even an Open API. From what I have seen, with the help of VC backing and FAANG endorsement, GraphQL mostly rose to replace JSON:API, which accomplishes pretty much all of the same goals in just JSON (and is RESTful).
One big issue of GraphQL is also that API clients tend to suck. That's not a problem for OpenAPIs.
And again, why is this the case? Some arbitrary design decision.
I feel like in general, someone creating a new DSL where it's not needed (and is obviously designed to look cool rather than actually be expressive), is a good sign they're just writing the software to stroke their ego rather than reach a meaningful end.
That's why in all the promo material for GraphQL you only see the query language, and not all of the actual setup required to send a request or even what an actual GraphQL HTTP request looks like. GraphQL, the framework, is not actually as simple and elegant as GraphQL the query language attempts to portray it as.
It's almost like someone came up with a query language for fun then came up with all the details of how a web server would utilize it afterwards.
Even today, GraphQL markets itself only as a query language (A query language for your API). When, as you have already mentioned, it is more than that.
That's why most developers know vaguely what GraphQL is ("Oh, that one language") but not how it actually works in practice. And when they actually encounter it, it feels almost like a betrayal, because it's nowhere near as simple, sleek or elegant as all the marketing they saw suggested.
At least, this was my experience when having to deal with a third party GraphQL API (funny enough, they migrated from REST, see ShipHero).
I usually tend to use words like these in sentences wherein there are phrases or clauses in which those words are the subject.
In all seriousness though, I mentioned JSON:API because it is a standard one can easily Google and it is a graph based data access API.
I wanted to demonstrate that it was possible for GraphQL to have been RESTful with the same ontology and mechanics.
I couldn't care less if someone created a JSON:API fork that handled relations like GraphQL, it would still be heavily preferred to GraphQL.
Also, the format for filtering in JSON:API is not defined by the spec. In other words, you can write whatever wacky query language you want.
When you create a stack of REST APIs, you’re creating a DSL. But it’s a DSL with arbitrary and frequently undocumented relationships between objects, no type safety, and masses of hidden complexity.
GraphQL is simple. If you don’t think it’s simple, you don’t understand it.
> One big issue of GraphQL is also that API clients tend to suck. That's not a problem for OpenAPIs.
The clients are unnecessary. You can get ridiculously complex clients for REST, too. But you can also use GraphQL just using fetch().
The only material difference between the two from a client perspective is:
* REST gives you everything, even if you don’t want it
* GraphQL requires you to request what you want using a minimal query language.
GraphQL also lets you perform multiple queries in parallel, on the server, which REST can’t easily do.
REST is a PITA for any data model that’s more complex than CRUD.
Also I'd just like to point out that for everything but queries themselves, GraphQL uses JSON. JSON is used in the C and U of GraphQL's CRUD... Explain to me why this couldn't have just been JSON and therefore a REST API again..?
Almost the only thing stopping it from being one is its "query language". I guess technically usage of one endpoint would make OpenAPI doc'ing difficult but I think possible depending on what JSON Schema/OpenAPI version you're using.
But it would also be trivial to just have separate endpoints for each schema.
You could even use all POST requests and only use the request body if you want.
Most clients send queries as json too, actually, not gql source. The funny thing is that GraphQL doesn't even specify JSON, or any wire format at all. It just has to be able to encode the gql type system, and JSON "just happens" to work (I'm pretty sure they did have json in mind when they wrote it, but it still isn't baked into the spec). The spec is also silent on transport, I think the whole tunneling-everything-through-POST thing came from the reference implementation.
You really seem attached to OpenAPI. I can't speak for everyone, but I for one would would much rather write SDL than json schema in yaml.
That's why I mentioned OpenAPIs, which you certainly can't make the same claims about. I have never had a problem with an OpenAPI, it's almost like a litmus test of web developer competence how simple it is.
> But instead of a complex mishmash if positional and named parameters, it has a super simple query structure.
Have you ever used an OpenAPI or an OpenAPI client? I seriously implore you to look at ShipHero's GraphQL documentation and look at any OpenAPI docs and pretend like GraphQL is simpler.
But if you've ever used an OpenAPI client, you know this is just a problem in theory, not in practice. I have never accidentally passed a parameter to the body, path, query or headers when it should've been elsewhere.
The problem is non-existent, especially since most people don't write OpenAPI requests from scratch.. you can just throw the document anywhere and you will have some easier way to make requests.
And even if that's really a problem, almost every OpenAPI doc UI I've seen has an option to show you the cURL and substitutes in parameters from the UI. You could just fill out the UI and copy the cURL request.
> When you create a stack of REST APIs, you’re creating a DSL. But it’s a DSL with arbitrary and frequently undocumented relationships between objects, no type safety, and masses of hidden complexity.
I'm beginning to think this entire comment is a strawman: my argument is that a graph-based OpenAPI would have been better than GraphQL. Any REST API can be an OpenAPI. You are comparing apples to oranges.
I am comparing a RESTful Graph based data access API (like JSON:API) to GraphQL.
You are taking what sounds like the kind of REST API you'd find in a w3schools tutorial and comparing it to GraphQL.
> GraphQL is simple. If you don’t think it’s simple, you don’t understand it.
That many developers find GraphQL difficult to understand is literally a testament to its complexity. This statement contradicts itself unless you believe I'm the only developer on the internet with this opinion, which I have a hard time believing given this entire comment is filled with strawman arguments you clearly use against others who dislike GraphQL.
> The clients are unnecessary. You can get ridiculously complex clients for REST, too. But you can also use GraphQL just using fetch().
Types are unnecessary. You can get ridiculously complex with strictly typed languages, too. But you can also just type everything dynamically and go crazy if the type system in my wacky language annoys you.
> REST gives you everything, even if you don’t want it
That's completely up to the REST API. I've written several that disagree with you. I've literally written RESTful Graph based data access APIs that don't do this. Multiple. At tech startups.
> GraphQL requires you to request what you want using a minimal query language.
So it requires complexity to enforce low bandwidth, got it. There is definitely no other way to lower bandwidth than enforcing complexity.
Usually I just use ABAC. You get the fields you can see. You can select a minimal list if you want. You don't have to type out 32 field names if you need them all.
I sometimes wish GraphQL just took a chapter from SQLs book. What you described here is not a feature.
> GraphQL also lets you perform multiple queries in parallel, on the server, which REST can’t easily do.
Ironically, that is up to the GraphQL implementation. ShipHero's does not support this, and requires N+1 queries in many common cases. I suspect that's not uncommon with GraphQL APIs, but I have no evidence for this.
But my suspicion is based on the fact that rigging up a GraphQL API implementation is a daunting enough task on its own, and multiple queries feels like it would be an after-thought to someone doing a quick-and-dirty setup. Which is how most people set up bleeding-edge infrastructure (which is what GraphQL was when it was largely popular), because there's not much knowledge, documentation or expertise readily available.
> REST is a PITA for any data model that’s more complex than CRUD.
Again, you're strawmanning. You're arguing against REST, not a RESTful Graph based data access API, like JSON:API.
REST is not an API. GraphQL is. It is basically an API proxy with a query language disguised as a graph based data access API.
Okay mr startup man, fact is you don’t know what you’re talking about and your head is so far up your startup ass you no longer know what’s good engineering.
Why are you writing your own restful graph based data access APIs?
The reality is that GraphQL is a great tool for one job: building APIs that scale across thousands of engineers, billions of users, and many megawatts of machines. This is what we use GraphQL for at Meta.
To make it work well, it takes dedicated teams building infrastructure that makes the DevX great. It takes many years to build this out across various backends, clients, and internal tools. We have open sourced some of this infra — the GraphQL spec, Relay, GraphiQL, etc.
But there are many tools we built that we have not open sourced yet, largely because they’re tightly coupled to our internal ORMs, runtimes, error handling, authentication and security systems, and so on. I wish we open sourced more of our GraphQL stack, so more people could experience the awesome DevX we have internally.
At the same time, GraphQL is a tool. No tool can be used all the time for everything. Meta’s 3rd party APIs still use REST for this reason; some internal services use Thrift; etc.
It’s up to you to evaluate the tools available to build the way that makes sense for your business. Because some parts are available off the shelf, but others often need to be built in house, it is up to you to understand that cost and evaluate whether it is worth it for your use cases.
My hope is that as time goes on, more dev tools startups come in the fill the gap in tooling around client libraries, auth, security, and so on to make GraphQL a good choice at smaller scales, too. The DevX is that good when you get it right.
Just like the title of the article suggests, GraphQL is great, but ONLY in the right context.
- Good support for HTTP 1.1
- Provides typesafe contracts for queries and mutations
- Strong support for most common UI clients (JS/React Native/SwiftUI)
- Backend can be written in a language that isn't JavaScript/TypeScript
- Large developer community
Closest thing I've found is defining your schema using OpenAPI spec, and using tools to generate code from that source for the client and server, but this is a much worse DX than the process of defining GraphQL schemas.
If you're coding resolvers by hand, that is not fun. But for example the Lighthouse GraphQL layer for Laravel (it's a schema-first thing based on the Webonyx PHP GraphQL reference implementation) can do a lot of this for you using schema decorations.
It has where/filter/sort/pagination, automatic scope narrowing etc., and you have very little code (if any) beyond the graphql schema to set it up. It integrates brilliantly with the Eloquent query builder.
One of the best applications for PHP IMO; it is really enormously productive.
GraphQL is just the protocol. It is up to implementations to add support for it. We use Hot Chocolate GraphQL server with entity framework backend, and it has built-in support for all of those things.
Vast majority of the criticism of GraphQL comes from people who do not understand it. GraphQL does have a steep learning curve.
GraphQL, but more specifically GraphQL Federation solves an organisational problem. With Federation, teams can collaborate on building an API together with a Schema Registry and Composition Checks across all Subgraphs (Services), which is a unique enabler in the market.
We've got a lot of customers in the enterprise segment and it's growing rapidly to the point where I'd say it's becoming an enterprise standard.
If you are building on Postgraphile and raising from angels, dm me.
GQL can abstract auth layers for external services, the user only has to auth with your service. The user only has to keep track of your single GQL url path instead of remembering “GET /users, PUT /posts, etc”. You’d probably have a premade GQL playground setup at the single URL too. If you forget the query options, do an introspection query, which most interactive GQL clients do in the background for you. It also makes caching/memos really easy, as you can memoize the GQL layer based on user query instead of implementing it for each upstream resource. Same deal with rate limiting. Also being able to ask for arbitrary number of resources by just making more named queries in a single request can be useful.
Having the .gql file be THE source of truth and typing for your schema is really nice. Most other setups I've seen like OpenAPI are usually disconnected from your actual API code, where it’s technically possible to diverge from your stated schema.
However I think we've seen a lot of 'GraphQL as the default paradigm from the get-go', which I think is a trap and has caused a fair bit of pain for all the small startups out there that were sold on this.
If a framework requires you to be just as, if not more competent in its niches to achieve the same end you could've achieved without it, I think that makes it bad.
That's like if an HTTP framework could only be effective at the protocol level with a team competent in HTTP. The whole point of a framework is abstraction.
I have never once seen how GraphQL (the language) provides any sort of abstraction that the general framework itself doesn't, and the little abstraction GraphQL the framework does provide you can mostly sum up by just generating an OpenAPI over a set of schemas, with a whole host of benefits and few trade offs (on the implementation side, which is where your 10x devs usually live anyways).
It's weird. It's like the main benefit of GraphQL is that there are groups of developers who understand it, but the biggest drawback is that it is needlessly complex, which again, seems to indicate GraphQL is bad. It's an abstraction that creates needless complexity (that you necessarily can't reverse because it's apart of the spec..)
I essentially only do that for mutations, and they are no more complex than writing POST handlers.
Essentially all the rest is schema-first and it's _so_ quick and simple.
If you do need them, as you said it's going to be complex either way.
His point about the runtime complexity of an API being entirely distinct from how the interface to it's code is exposed (whether GraphQL or REST or otherwise) is fairly obvious, I think.
The counter-argument is that unlimited query complexity makes it far bigger problem, and the author's point is that if you're using it for private APIs with persisted queries, you shouldn't have that problem unknowingly.
Don't get me wrong - I think the takeaway is that GraphQL's niche is quite small, and he's defending exactly that niche. It's not often the case that you can develop an API in a private manner which doesn't undercut higher-order value in the future, as the rise of AWS hopefully made evident.
Though I guess the layer in between lets you change your SQL schema without changing the GraphQL schema.
SQL has got views for that.